shader_recompiler: Implement S_BCNT1_I32_B64

This commit is contained in:
squidbus 2024-12-25 21:33:46 -08:00
parent 5f9abd1db4
commit eeb732572a
7 changed files with 25 additions and 3 deletions

View File

@ -304,6 +304,7 @@ Id EmitBitFieldSExtract(EmitContext& ctx, IR::Inst* inst, Id base, Id offset, Id
Id EmitBitFieldUExtract(EmitContext& ctx, IR::Inst* inst, Id base, Id offset, Id count);
Id EmitBitReverse32(EmitContext& ctx, Id value);
Id EmitBitCount32(EmitContext& ctx, Id value);
Id EmitBitCount64(EmitContext& ctx, Id value);
Id EmitBitwiseNot32(EmitContext& ctx, Id value);
Id EmitFindSMsb32(EmitContext& ctx, Id value);
Id EmitFindUMsb32(EmitContext& ctx, Id value);

View File

@ -201,6 +201,10 @@ Id EmitBitCount32(EmitContext& ctx, Id value) {
return ctx.OpBitCount(ctx.U32[1], value);
}
Id EmitBitCount64(EmitContext& ctx, Id value) {
return ctx.OpBitCount(ctx.U64, value);
}
Id EmitBitwiseNot32(EmitContext& ctx, Id value) {
return ctx.OpNot(ctx.U32[1], value);
}

View File

@ -100,6 +100,8 @@ void Translator::EmitScalarAlu(const GcnInst& inst) {
return S_BREV_B32(inst);
case Opcode::S_BCNT1_I32_B32:
return S_BCNT1_I32_B32(inst);
case Opcode::S_BCNT1_I32_B64:
return S_BCNT1_I32_B64(inst);
case Opcode::S_FF1_I32_B32:
return S_FF1_I32_B32(inst);
case Opcode::S_AND_SAVEEXEC_B64:
@ -585,6 +587,12 @@ void Translator::S_BCNT1_I32_B32(const GcnInst& inst) {
ir.SetScc(ir.INotEqual(result, ir.Imm32(0)));
}
void Translator::S_BCNT1_I32_B64(const GcnInst& inst) {
const IR::U32 result = ir.BitCount(GetSrc64(inst.src[0]));
SetDst(inst.dst[0], result);
ir.SetScc(ir.INotEqual(result, ir.Imm32(0)));
}
void Translator::S_FF1_I32_B32(const GcnInst& inst) {
const IR::U32 src0{GetSrc(inst.src[0])};
const IR::U32 result{ir.Select(ir.IEqual(src0, ir.Imm32(0U)), ir.Imm32(-1), ir.FindILsb(src0))};

View File

@ -111,6 +111,7 @@ public:
void S_NOT_B64(const GcnInst& inst);
void S_BREV_B32(const GcnInst& inst);
void S_BCNT1_I32_B32(const GcnInst& inst);
void S_BCNT1_I32_B64(const GcnInst& inst);
void S_FF1_I32_B32(const GcnInst& inst);
void S_GETPC_B64(u32 pc, const GcnInst& inst);
void S_SAVEEXEC_B64(NegateMode negate, bool is_or, const GcnInst& inst);

View File

@ -1273,8 +1273,15 @@ U32 IREmitter::BitReverse(const U32& value) {
return Inst<U32>(Opcode::BitReverse32, value);
}
U32 IREmitter::BitCount(const U32& value) {
return Inst<U32>(Opcode::BitCount32, value);
U32 IREmitter::BitCount(const U32U64& value) {
switch (value.Type()) {
case Type::U32:
return Inst<U32>(Opcode::BitCount32, value);
case Type::U64:
return Inst<U32>(Opcode::BitCount64, value);
default:
ThrowInvalidType(value.Type());
}
}
U32 IREmitter::BitwiseNot(const U32& value) {

View File

@ -229,7 +229,7 @@ public:
[[nodiscard]] U32 BitFieldExtract(const U32& base, const U32& offset, const U32& count,
bool is_signed = false);
[[nodiscard]] U32 BitReverse(const U32& value);
[[nodiscard]] U32 BitCount(const U32& value);
[[nodiscard]] U32 BitCount(const U32U64& value);
[[nodiscard]] U32 BitwiseNot(const U32& value);
[[nodiscard]] U32 FindSMsb(const U32& value);

View File

@ -284,6 +284,7 @@ OPCODE(BitFieldSExtract, U32, U32,
OPCODE(BitFieldUExtract, U32, U32, U32, U32, )
OPCODE(BitReverse32, U32, U32, )
OPCODE(BitCount32, U32, U32, )
OPCODE(BitCount64, U32, U64, )
OPCODE(BitwiseNot32, U32, U32, )
OPCODE(FindSMsb32, U32, U32, )